cannot reproduce the result in the paper
zhoudayang opened this issue · comments
the photo showed that the performance of reside is worse than pcnn
Seems like the model has not converged, you need to train it for more epochs with a lower learning rate. After 3-4 epochs with Adam you can train it further with SGD with learning rate .001
or lower. I think then it should be fine. You can do that easily by executing:
python -restore -name <full_name_of_run> -opt sgd -lr 0.001
thanks, I will try it again asap.
Hi @zhoudayang.
I looked into the problem, there was a slight error in copying hyperparameters. You can pull the code and start afresh training and let it get finished on its own.
python reside.py -name new_train_run -data data/riedel_processed.pkl
After that, you can restore and train the model again with SGD for 2-3 epochs
python reside.py -name new_train_run_xx_xx_xxxx_xx:xx:xx -restore -opt sgd -lr 0.001 -l2 0.0 -epoch 3
Then run python plot_pr.py -name new_train_run_20_11_2018_13:05:18
to get the plot.